IC-LoRA Ingredients (Reference Sheet)
Lay out your cast, props and location on a single image and LTX-2.5 will compose a scene from them. In the prompt, refer to each element by where it sits on the sheet ("the woman top-left", "the red car bottom-right"). Useful for keeping characters consistent across shots.
Required models
Drop each file into the matching folder under ComfyUI/models/, or let Workflow Overview download them all.
| File | Folder | Size | Purpose | Link |
|---|---|---|---|---|
ltx-2.5-22b-distilled-transformer-bf16.safetensors | ComfyUI/models/diffusion_models/ | 42 GB | The LTX-2.5 22B distilled video model — used by every example workflow. Few sampling steps, built for fast iteration. | Hugging Face |
↳ lower-VRAM alternative ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors | ComfyUI/models/diffusion_models/ | 21.5 GB | INT8 — about half the size; the usual pick for 24 GB cards. | Hugging Face |
↳ lower-VRAM alternative ltx-2.5-22b-distilled-transformer-nvfp4.safetensors | ComfyUI/models/diffusion_models/ | 18.7 GB | NVFP4 — smallest; needs an NVIDIA GPU with FP4 support (Blackwell). | Hugging Face |
gemma4-12b-with-proj-ltx-2.5-bf16.safetensors | ComfyUI/models/text_encoders/ | 26.3 GB | Gemma 4 12B text encoder with the LTX-2.5 projection — turns your prompt into conditioning. Required by all workflows. | Hugging Face |
↳ lower-VRAM alternative gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors | ComfyUI/models/text_encoders/ | 15.4 GB | INT8 text encoder — saves ~11 GB on disk and in memory. | Hugging Face |
ltx-2.5-video-vae-bf16.safetensors | ComfyUI/models/vae/ | 1.5 GB | Video VAE — encodes input frames and decodes the generated latents to pixels. | Hugging Face |
ltx-2.5-video-vae-conv-bf16.safetensors | ComfyUI/models/vae/ | 1.5 GB | Convolutional video VAE variant referenced by the example graphs' decode path. | Hugging Face |
ltx-2.5-audio-vae-bf16.safetensors | ComfyUI/models/vae/ | 0.4 GB | Audio VAE — decodes the synchronized soundtrack LTX-2.5 generates alongside the video. | Hugging Face |
ltx-2.3-22b-ic-lora-ingredients-0.9.safetensors | ComfyUI/models/loras/ | — | IC-LoRA: generate from a reference sheet of characters, props and locations. | Hugging Face |
How to use
- 1
Open ComfyUI (current version, with ComfyUI-LTXVideo installed via Manager).
- 2
Workflow → Open and pick the downloaded .json, or drag the file onto the canvas.
- 3
Open Workflow Overview → Missing Models → Download all, and wait for the files to finish.
- 4
Assemble a reference sheet: characters, props, wardrobe, location in a clear grid.
- 5
Load it as the IC-LoRA guide image.
- 6
Prompt the scene, naming elements by position on the sheet, then Run.
VRAM
Which transformer build fits your card for this graph.
| VRAM | Example cards | Verdict | Transformer | Advice |
|---|---|---|---|---|
| 24 GB | RTX 3090, 4090, A5000 | OK | int8 | Workable with the INT8 transformer + INT8 text encoder and the low-VRAM loader nodes. Two-stage graphs run; keep decode tiles small. Leave bf16 files for 32 GB+ cards. |
| 32 GB | RTX 5090, V100 32G | Comfortable | bf16 | Lightricks' stated minimum. bf16 transformer with the low-VRAM loaders fits; INT8 text encoder frees headroom for larger decode tiles and longer clips. |
| 48 GB+ | RTX 6000 Ada, A6000, H100 | Comfortable | bf16 | Everything in bf16 without offloading. Increase the decode tile size (see the Decode notes in each graph) for faster runs. |
Official requirement is 32 GB+. Lower tiers are guidance for quantized builds.
Under 24 GB?
The same LTX-2 model runs in our browser generator — 4K with audio, no setup, free credits to start.
Troubleshooting
FAQ
Other workflows
IC-LoRA Union Control (Depth / Canny / Pose)
Video-to-video driven by a depth, canny or pose annotator run on a reference clip. Depth is wired by default.
Video to Video — IC-LoRA Edit
Use the source frames as IC-LoRA guides and change appearance with an identity / style / edit LoRA. Original audio is kept.
IC-LoRA Motion Track
Image-to-video with sparse motion tracks you draw on the first frame. Direct exactly how objects move.